Stella Laurenzo comments

Results 204 comments of


                                            Stella Laurenzo

First step of adding nccl headers to the build.

Yeah, I was figuring it needed to be optional at some point which is why I have the c guard -- it is just unconditionally true now and could use...

First step of adding nccl headers to the build.

No, I was just going to dlopen the system one at runtime. I haven't had my coffee yet either so may be missing something, but this seems all set up...

First step of adding nccl headers to the build.

Sorry - I could have mentioned that. I was thinking the better was may be too just let external_project pull it in. The pinning is more so that builds don't...

First step of adding nccl headers to the build.

I was debating that point myself: convinced. I'll keep the cuda configuration goo on this patch as it is an improvement over what we had before, which I couldn't get...

First step of adding nccl headers to the build.

We would just add the handful of enum/values inline with the function declarations that use them. Given the PyTorch precedent, this is pretty straight-forward and the simplicity of this library...

First step of adding nccl headers to the build.

> Hey @stellaraccident, the final piece of #8971 is to enable CUDA by default. > > Given this change, iis still relevant since this is changing the behavior of CUdA...

[spirv] PyTorch SqueezeNet returns NaN values on Apple M GPUs

Does this only repro on the M2? I don't think that anyone in the iree core team has that hardware.

Remove Config() constructor in favor of explicit API for creating devices

@krzys-ostrowski

Remove Config() constructor in favor of explicit API for creating devices

Yes, i have a patch for this but needs reviving

Support device side buffers for Python API

Done as part of the DeviceArray work.